Anthropic Launches Project Deal: Claude Agents Complete 186 Autonomous Financial Transactions

Anthropic has released the results of "Project Deal," an ambitious experiment testing the ability of AI agents to handle real-world economic negotiations. In a closed-loop marketplace involving nearly 70 participants, Claude-based agents autonomously negotiated, listed, and purchased items, completing 186 successful transactions totaling over $4,000.

The research revealed a significant performance gap between model tiers. Agents powered by Claude 3.5 Opus consistently outperformed those using the Haiku model, securing better prices for their human owners. Surprisingly, participants represented by the less capable models often didn’t realize they were getting worse deals, highlighting a new form of "Economic AI Asymmetry" that could define future marketplaces.

Anthropic’s team observed instances of social coordination and even humorous agent behavior, such as one agent negotiating like an "exasperated cowboy." The project demonstrates that the technology for "Agentic Commerce" is no longer theoretical, paving the way for a future where personal AI assistants manage our subscriptions, purchases, and complex financial negotiations with minimal human oversight.